Semisupervised white matter hyperintensities segmentation on MRI.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Cerebral white matter hyperintensities (WMHs) are crucial imaging biomarkers for various neurological conditions.
• Accurate segmentation of WMHs is essential for diagnosis and monitoring disease progression.
• Current segmentation methods often rely on extensive manual labeling, which is time-consuming and labor-intensive.

Purpose of the Study:

• To propose a semisupervised loss function, Level-Set Loss (LSLoss), for automated WMH segmentation.
• To develop and evaluate a V-Net model trained with LSLoss on diverse MRI datasets.
• To demonstrate the efficacy of the semisupervised approach in segmenting WMHs without requiring manually labeled masks.

Main Methods:

• Implemented a semisupervised learning framework utilizing a novel LSLoss function.
• Preprocessed MRI data including bias field correction, skull stripping, and white matter segmentation.
• Trained a V-Net model on a combined dataset from local (HKU-SVD, HKU-MS) and public (MICCAI-WMH, ADNI-CN) databases.

Main Results:

• Achieved high Dice Similarity Coefficients (DSC): 0.81 on HKU-SVD, 0.77 on HKU-MS, and 0.78 on MICCAI-WMH testing sets.
• Demonstrated segmentation performance comparable to supervised methods.
• Outperformed existing unsupervised segmentation methods in the literature.

Conclusions:

• The proposed LSLoss enables accurate and efficient semisupervised segmentation of WMHs.
• This method reduces the dependency on manual annotations, making WMH segmentation more accessible.
• The V-Net model with LSLoss shows significant potential for clinical application in neurological disease assessment.